Brain Function Connection Analysis Of Resting State FMRI Based On Machine Learning

Functional magnetic resonance imaging(fMRI)has been widely applied in brain disease diagnosis due to its advantages of high spatial and temporal resolution,non-invasive,and so on.The Function Connection Network(FCN)built based on fMRI helps find biomarkers and pathogenic factors related to brain diseases.This study is based on FCN analysis of resting state fMRI images of mild traumatic brain injury(mTBI)to break through the main bottlenecks in diagnosing and treating mTBI:lack of biomarkers and difficulty in identifying pathological mechanisms.With the development of artificial intelligence,machine learning methods provide new tools and ideas for fMRI analysis.However,the brain network data of mTBI patients face the problem of small samples and high dimensions,which brings difficulties to diagnosis based on machine learning methods.This study aims to extract the most discriminating functional connectivity features from high-dimensional and complex FCNs,further explore relevant biomarkers,and analyze the pathological mechanism of mTBI.The following two aspects of work have been completed:This paper designs a hierarchical feature selection pipeline(HFSP)composed of Variance Filtering(VF),Lasso,and Principal Component Analysis(PCA)in sequence.Firstly,feature variance reflects the importance of features for classification,and the greater the variance,the greater the feature discrimination.Therefore,the lightweight VF method is adopted to filter redundant features whose variance is less than a threshold value.However,there is still a strong correlation between the remaining features.Lasso further eliminated many redundant strong correlation features.Finally,PCA is utilized to reduce the dimensions to obtain the fused principal components,which are input into the classifier for classification verification.To test the impact of each feature selection module on classification,this study designed ablation experiments to verify the robustness and reliability of HFSP.In addition,this paper conducted comparative experiments from the perspectives of feature selection and classifier.For the effect of feature selection,the HFSP is compared with recursive feature elimination,elastic networks,and locally linear embedding,and the results show that HFSP has significant advantages.To verify the classification performance,the HFSP is combined respectively with random forest,SVM,Bayesian,linear discriminant analysis,and logical regression to verify the universality of HFSP.This paper identified 25 pairs of the most discriminating functional connections through connectome analysis to explore mTBI-related biomarkers,mainly concentrated in the frontal lobe,occipital lobe,and cerebellum.These 25 pairs of functional connectivity visualizations show nine brain regions with the highest nodal degree,corresponding to the mTBI-damaged brain regions.On this basis,this paper conducted a brain network analysis from a macro perspective to correlate the pathological mechanism of mTBI injury.The results showed that the discriminating functional connections in patients with mTBI were mainly distributed in the cerebellar network,visual network,and somatic motor network.To further reveal the causal relationship between various damaged brain regions of mTBI,this paper conducted a Granger Causality Analysis,providing more neuroimaging evidence for clarifying the pathogenesis of mTBI.The results confirm extensive coordination and interaction between the cerebellum and other network nodes.More attention should be paid to cerebellar dysfunction in the precise treatment and follow-up of mTBI.In summary,discriminating functional connections in mTBI affect brain network connectivity and function through whole-brain distribution and local aggregation,leading to motor function,executive function,and cognitive and emotional disorders in patients with mTBI.